www.nature.com/scientificreports OPEN MAGPEL: an autoMated pipeline for inferring vAriant‑driven Gene PanEls from the full‑length biomedical literature Nafseh Saberian1, Adib Shaf 1, Azam Peyvandipour1 & Sorin Draghici 1,2* In spite of the eforts in developing and maintaining accurate variant databases, a large number of disease‑associated variants are still hidden in the biomedical literature. Curation of the biomedical literature in an efort to extract this information is a challenging task due to: (i) the complexity of natural language processing, (ii) inconsistent use of standard recommendations for variant description, and (iii) the lack of clarity and consistency in describing the variant-genotype-phenotype associations in the biomedical literature. In this article, we employ text mining and word cloud analysis techniques to address these challenges. The proposed framework extracts the variant- gene‑disease associations from the full‑length biomedical literature and designs an evidence‑based variant-driven gene panel for a given condition. We validate the identifed genes by showing their diagnostic abilities to predict the patients’ clinical outcome on several independent validation cohorts. As representative examples, we present our results for acute myeloid leukemia (AML), breast cancer and prostate cancer. We compare these panels with other variant‑driven gene panels obtained from Clinvar, Mastermind and others from literature, as well as with a panel identifed with a classical diferentially expressed genes (DEGs) approach. The results show that the panels obtained by the proposed framework yield better results than the other gene panels currently available in the literature. One crucial step in understanding the biological mechanism underlying a disease condition is to capture the relationship between the variants and the disease risk1. Tere are several publicly available databases contain the disease-associated variants such as Clinvar2, SNPedia3, OMIM4, Swiss-Prot5, COSMIC6, BioMuta7, HGMD8, UMD9, HGVbaseG2P10, MutDB11, dbSNP12, PharmGKB13 and InSiGHT14. All these databases are manually curated by human experts. While this manual curation ensures a high quality of the annotations, the manual extraction of this type of information from the biomedical literature takes an enormous amount of time and efort. Te current rate with which new variants are published is simply too high for any manual annotation process. As an additional challenge, despite the HGVS (Human Genome Variation Society) standard recommen- dations for the description of the variants, many variants are still reported in literature in non-standard formats. A number of automatic mutation indexing tools have been developed. Such tools process biomedical literature and produce a list of mutations that appear in these papers. Tese include MutationFinder 15, EMU16, MEMA17, MuteXt18, Mutation GraB 19 and MutationMiner20. Te most recent such tool, tmVar 2.021 extracts variants from an article and normalizes them to their unique dbSNP identifers. Te next step is to develop sofware tools to extract variants-disease associations from the biomedical literature. Several methods have been proposed for this purpose such as MuGeX22, OSIRIS23, EnzyMiner24 and the methods proposed by Singhal et al.1,25. All these methods have been applied to only the title and the abstract section of biomedical articles. However, a com- prehensive study showed that a signifcant number of genetic variants are only included in the full text and the supplementary materials of the articles 26. Tese will be missed if the variants are only extracted from titles and abstracts. Doughty et al.16 also proposed a tool named EMU for extracting the disease-associated mutations from biomedical literature. Although this tool automatically extracts the mutations and their corresponding genes from an article, it still requires human curation to discover the mutation-disease associations. 1Department of Computer Science, Wayne State University, Detroit, MI, USA. 2Department of Obstetrics and Gynecology, Wayne State University, Detroit, MI, USA. *email: [email protected] SCIENTIFIC REPORTS | (2020) 10:12365 | https://doi.org/10.1038/s41598-020-68649-0 1 Vol.:(0123456789) www.nature.com/scientificreports/ Figure 1. Framework overview. Module (A) obtains all the publicly available full-length articles from the PubMed Central (PMC) database. Ten it uses the word cloud analysis and generate a weighted list of variant- relevant keywords. Te variant-relevant articles are then selected based on the presence of this list in their full text (“Variant-relevant input corpus”). Module (B) uses GNormPlus27, tmVar 2.021 and DNorm28 tools to extract the gene, variant and disease phenotype entities, respectively (“Extract the variant, gene and disease entities”). Module (C) extracts the gene-variant associations from each input article (“Extract the variant-gene associations”). Tis module also uses a set of features to discover the disease-variant associations (“Extract the variant-disease associations”). Module (D) generates a panel consists of the variant-gene-disease associations. Here we propose an automated framework to extract disease-associated variants from the full-length biomedi- cal literature and design a variant-driven gene panel for a given disease phenotype. As the frst step, the proposed framework employs word cloud analysis to identify the variant-relevant articles. Te variant-gene-disease asso- ciations are then extracted from these articles. An evidence-based variant-driven gene panel is then generated based on the mined triplet information. A comprehensive validation procedure illustrates the capabilities of the proposed framework. We validate the proposed variant-driven gene panels by showing their abilities to predict the patients’ clinical conditions (healthy vs. disease) on multiple independent validation datasets. Methods Figure 1 illustrates the proposed framework that consists of the following four major modules: (1) obtain the full-length variant-relevant articles; (2) extract all the variant, gene and disease entities from each input article; (3) identify the variant-gene and the variant-disease associations in each input article; (4) design a variant-driven gene panel for a given phenotype. Te detailed descriptions of each step are provided in the following sections. Variant‑relevant input corpus. Te input of the proposed framework consists of 3,322,746 full-length articles downloaded from the PMC database on January 2020. Te variant indexing procedure from a full-length article is challenging because any chemical formulae, fgure numbers, etc. that are represented in “Character- Number-Character” format could be identifed as a variant21. One solution to address this challenge is to mine only the variant-relevant articles. We compare the performances of two diferent approaches for detecting the variant-relevant articles. Te frst approach considers only the articles that mention any disease or gene or any of their synonyms in the title and abstract sections. In the second approach, we employ the word cloud analysis and generate a weighted list of variant-relevant keywords. In particular, we frst generate a weighted list of words (referred to as variant-relevant keywords) that appear frequently in the full-body text of 10,000 random articles SCIENTIFIC REPORTS | (2020) 10:12365 | https://doi.org/10.1038/s41598-020-68649-0 2 Vol:.(1234567890) www.nature.com/scientificreports/ Figure 2. Among the 10,000 random articles, the articles with at least one mentioned mutation are selected (using tmVar 2.0). We compare the performances of two diferent approaches for detecting the variant-relevant articles. Te frst approach identifes articles that mention any disease or gene or any of their synonym in their titles and abstracts. In the second approach, we only search for the articles that mention the variant- relevant keywords in their full-body text. Te variant-relevant keywords is a weighted list of the words that appear frequently in a set of 10,000 random articles with at least one mentioned variants (using tmVar 2.0). Subsequently, an article is considered to be relevant to variants if at least 10% of these variant-relevant keywords are appear in the full-body text. Te number of variants that are found in the articles selected by the frst approach and the second approach are 5,760 and 6,087, respectively. Te number of variants identifed by both approaches is 5,476. Te number of variants that are only found by the frst approach is 284, of which 97% are false positive (unrelated text wrongly identifed as a variant). Te number of variants that are only found by the second approach is 611, of which only 10% are false positive. Tese results show that the second approach which is based on the variant-relevant keywords outperforms the frst approach. with at least one mentioned variant (using tmVar 2.0). Subsequently, an article is considered to be relevant to variants if at least 10% of these keywords appear in the full-body of the article. We apply both approaches on a set of 10,000 random full-length articles. Figure 2 shows the identifed variant overlaps and diferences between the two approaches. Te number of papers with at least one mentioned variant overlapped between the two approaches is 836 and the number of overlapped variants is 5,476. Te number of variants that are only found by the frst approach is 284 from 91 papers, in which a manual validation process revealed that 97% of them are false positive (unrelated text wrongly identifed as a variant). Te number of variants that are only found by the second approach is 611 from 122 papers, in which only 10% of them are false positive. Tese results show that the second approach which is based on the variant-relevant keywords outperforms the frst approach. Te manual validation of the extracted variants are included in the Supplementary materials (Table S1). Tis leads us to the conclusion that the second approach performs better in terms of the ability to index the variant-relevant articles. Tis approach results in a list of 1,274,775 full-length articles that contain genomic variants.